diff options
Diffstat (limited to 'arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h')
-rw-r--r-- | arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h | 92 |
1 files changed, 61 insertions, 31 deletions
diff --git a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h index 95261929ca..78ac56418f 100644 --- a/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h +++ b/arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h @@ -36,21 +36,35 @@ class NEWinogradLayerKernel; class Winograd3x3F32 final { public: + /** Create a new Winograd convolution layer. + * + * @param[in] n_batches Number of batches in the input and output tensors. + * @param[in] n_input_channels Number of feature maps in a batch of the input tensor. + * @param[in] n_input_rows Number of rows in a feature map of the input tensor. + * @param[in] n_input_cols Number of columns in a feature map of the input tensor. + * @param[in] n_output_channels Number of feature maps in the output tensor. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". + * @param[in] weights Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. + * @param[out] weights_storage Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size + * @param[in] input Pointer to NHWC ordered input tensor, in the spatial domain. + * @param[out] winograd_input Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. + * @param[out] output Pointer to NHWC ordered output tensor, in the spatial domain. + * @param[out] winograd_output Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. + */ friend class NEWinogradLayerKernel; Winograd3x3F32( - const int n_batches, /** Number of batches in the input and output tensors. */ - const int n_input_channels, /** Number of feature maps in a batch of the input tensor. */ - const int n_input_rows, /** Number of rows in a feature map of the input tensor. */ - const int n_input_cols, /** Number of columns in a feature map of the input tensor. */ - const int n_output_channels, /** Number of feature maps in the output tensor. */ - const bool same_padding, /** Use "SAME" padding, otherwise use "VALID". */ - const float *const weights, /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */ - float *const weights_storage, /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */ - const float *const input, /** Pointer to NHWC ordered input tensor, in the spatial domain. */ - float *const winograd_input, /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */ - float *const output, /** Pointer to NHWC ordered output tensor, in the spatial domain. */ - float *const winograd_output /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */ - ); + const int n_batches, + const int n_input_channels, + const int n_input_rows, + const int n_input_cols, + const int n_output_channels, + const bool same_padding, + const float *const weights, + float *const weights_storage, + const float *const input, + float *const winograd_input, + float *const output, + float *const winograd_output); ~Winograd3x3F32(); void transform_weights(); @@ -88,31 +102,47 @@ public: // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; - /* Get the memory required to instantiate a new Winograd operator. - */ - static size_t get_weight_storage_size( - const int n_output_channels, /** Number of output feature maps. */ - const int n_input_channels /** Number of input feature maps. */ - ); + /** Determine how much memory (in units of TIn) to allocate for the + * transformed weights. + * + * @param[in] n_output_channels Number of output feature maps. + * @param[in] n_input_channels Number of input feature maps. + */ + static unsigned int get_weight_storage_size( + const int n_output_channels, + const int n_input_channels); + /** Determine how much memory (in units of TIn) to allocate for the + * transformed input. + * + * @param[in] n_batches Number of batches in the input tensor. + * @param[in] n_channels Number of feature maps in the input tensor. + * @param[in] n_rows Number of rows in each feature map. + * @param[in] n_cols Number of columns in each feature map. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". + */ static unsigned int get_input_storage_size( - const int n_batches, /** Number of batches in the input tensor. */ - const int n_channels, /** Number of feature maps in the input tensor. */ - const int n_rows, /** Number of rows in each feature map. */ - const int n_cols, /** Number of columns in each feature map. */ - const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ - ); + const int n_batches, + const int n_channels, + const int n_rows, + const int n_cols, + const bool same_padding); /** Determine how much memory (in units of TOut) to allocate for the * (Winograd domain) output. + * + * @param[in] n_batches Number of batches in the output tensor. + * @param[in] n_rows Number of rows in each feature map of the input tensor. + * @param[in] n_cols Number of columns in each feature map of the input tensor. + * @param[in] n_output_channels Number of feature maps in the output tensor. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". */ static unsigned int get_output_storage_size( - const int n_batches, /** Number of batches in the output tensor. */ - const int n_rows, /** Number of rows in each feature map of the input tensor. */ - const int n_cols, /** Number of columns in each feature map of the input tensor. */ - const int n_output_channels, /** Number of feature maps in the output tensor. */ - const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ - ); + const int n_batches, + const int n_rows, + const int n_cols, + const int n_output_channels, + const bool same_padding); protected: Winograd3x3F32 *_convolver; |